Summary of The Pitfalls Of Memorization: When Memorization Hurts Generalization, by Reza Bayat et al.
The Pitfalls of Memorization: When Memorization Hurts Generalization
by Reza Bayat, Mohammad Pezeshki, Elvis Dohmatob, David Lopez-Paz, Pascal Vincent
First submitted to arxiv on: 10 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed paper formalizes the interplay between memorization and generalization in neural networks, demonstrating how spurious correlations can lead to poor generalization when combined with memorization. The authors show that memorization can reduce training loss to zero, eliminating the incentive for models to learn robust patterns. To address this issue, they introduce memorization-aware training (MAT), which uses held-out predictions as a signal of memorization to shift model logits and promote learning invariant patterns across distributions. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper investigates how neural networks learn simple explanations that fit most data while memorizing exceptions. This “this http URL” behavior can lead to poor generalization when learned explanations rely on spurious correlations. The authors show that memorization can reduce training loss, making it harder for models to learn robust patterns. To fix this, they propose MAT, which uses predictions from held-out data to help models focus on learning robust patterns that work across different situations. |
Keywords
» Artificial intelligence » Generalization » Logits
